Generative Adversarial Networks Using Adaptive Convolution

Most existing GAN architectures that generate images use transposed convolution or resize-convolution as their upsampling algorithm from lower to higher resolution feature maps in the generator. We propose a novel adaptive convolution method that learns the upsampling algorithm based on the local context at each location to address this problem. We modify a baseline GAN architecture by replacing normal convolutions with adaptive convolutions in the generator. Our method is orthogonal to others that seek to improve GAN by incorporating high level information. Experiments on CIFAR-10 dataset show that our modified models improve the baseline model by a large margin on visually diverse datasets.

I. Introduction

Generative Adversarial Networks [1] (GAN) are an unsupervised learning method that is able to generate realistic looking images from noise. GAN employs a minimax game where a generator network learns to generate synthesized data from random noise and in conjunction, a discriminator network learns to distinguish between real and generated data. Theoretically, the training processes of the two networks intertwine and iterate until both networks reach a Nash equilibrium where real and synthesized data are indistinguishable.